Download Math 210C. Weyl groups and character lattices 1. Introduction Let G

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Lie algebra extension wikipedia , lookup

Transcript
Math 210C. Weyl groups and character lattices
1. Introduction
Let G be a connected compact Lie group, and S a torus in G (not necessarily maximal). In
class we saw that the centralizer ZG (S) of S in G is connected conditional on Weyl’s general
conjugacy theorem for maximal tori in compact connected Lie groups. The normalizer NG (S)
is a closed subgroup of G that contains S. (Read the discussion just below Corollary 4.6 in
the handout on local/global Frobenius theorems for some subtleties concerning normalizer
subgroups of closed subgroups that are connected or have infinite component group.)
For any compact Lie group H, the component group π0 (H) = H/H 0 is finite since it is
compact and discrete. (as H 0 is open in H, due to the local connectedness of manifolds)
Thus, if H is a compact Lie group then a connected closed subgroup H 0 exhausts the identity
component H 0 precisely when H/H 0 is finite. (Indeed, in such cases H 0 /H 0 with its quotient
topology is a subspace of the Hausdorff quotient H/H 0 that is finite and hence discrete.
Thus, H 0 is open in H, yet also visibly closed, so its finitely many cosets contained in the
compact connected H 0 constitute a pairwise disjoint open cover, contradicting connectedness
of H 0 unless the cover is a singleton; i.e., H 0 = H 0 .)
Note that since ZG (S) is normal in NG (S) (why?), its identity component ZG (S)0 (which
we don’t know is equal to ZG (S) until the Conjugacy Theorem is proved) is also normal in
NG (S). In particular, NG (S)/ZG (S)0 makes sense as a compact Lie group. In this handout,
we will show that W (G, S) := NG (S)/ZG (S) is finite, which is equivalent to the finiteness
of the compact Lie group NG (S)/ZG (S)0 since π0 (ZG (S)) is finite. This finiteness holds if
and only if ZG (S)0 = NG (S)0 (why?). The group W (G, S) is the Weyl group of G relative
to S, and its primary interest is when S is a maximal torus in G (in which case ZG (S) = S
conditional on the Conjugacy Theorem, as we saw in class).
We will first look at some instructive examples when S is maximal (by far the most
important case for the study of connected compact Lie groups, for which G-conjugation
permutes all choices of such a torus, conditional on the Conjugacy Theorem). Then we will
prove two results: the finiteness of W (G, S) and that T = ZG (T )0 when T is maximal (so
NG (T )/T is also finite for maximal T , without needing to know that ZG (T ) = T ). These
two proofs will have no logical dependence on the Conjugacy Theorem, an important point
because the proof of the Conjugacy Theorem uses the finiteness of NG (T )/T for maximal T .
2. Examples
Let G = U(n) ⊂ GLn (C) with n > 1 and let T = (S 1 )n be the diagonal maximal torus in
G. There are some evident elements in NG (T ) outside T , namely the standard permutation
matrices that constitute an Sn inside G. We first show:
Lemma 2.1. The centralizer ZG (T ) is equal to T and the subgroup Sn of G maps isomorphically onto W (G, T ) = NG (T )/T .
Proof. We shall work “externally” by appealing to the vector space Cn on which G = U(n)
naturally acts. The diagonal torus T = (S 1 )n acts linearly on Cn in this way, with the
standard basis lines Cej supporting the pairwise distinct characters χj : t 7→ tj ∈ C× .
1
2
This gives an intrinsic characterization of these lines via the completely reducible finitedimensional C-linear representation theory of T . Thus, conjugation on T by any g ∈ NG (T )
must permute these lines! Likewise, an element of NG (T ) that centralizes T must act on
Cn in a manner that preserves each of these lines, and so is diagonal in GLn (C). In other
words, ZG (T ) is the diagonal subgroup of the compact subgroup U (n) ⊂ GLn (C). But by
compactness we therefore have ZG (T ) = T , since entry-by-entry we can apply the evident
×
1
fact that inside C× = S 1 ×R×
>0 any compact subgroup lies inside S (as R>0 has no nontrivial
compact subgroups).
Returning to an element g ∈ NG (T ), whatever its permutation effect on the standard
basis lines may be, there is visibly a permutation matrix s ∈ Sn ⊂ G that achieves the same
effect, so s−1 g preserves each of the standard basis lines, which is to say s−1 g is diagonal as
an element in the ambient GLn (C) and so it lies in ZG (T ) = T . Hence, g ∈ sT , so every class
in W (G, T ) is represented by an element of Sn . It follows that Sn → W (G, T ) is surjective.
For injectivity, we just have to note that a nontrivial permutation matrix must move some
of the standard basis lines in Cn , so its conjugation effect on T = (S 1 )n is to permute the S 1 factors accordingly. That is a nontrivial automorphism of T since n > 1, so Sn → W (G, T )
is indeed injective.
The preceding lemma has an additional wrinkle in its analogue for G0 = SU(n) in place
of G = U(n), as follows. In G0 , consider the diagonal torus T 0 ofQdimension n − 1. This
is the kernel of the map (S 1 )n → S 1 defined by (z1 , . . . , zn ) 7→
zj , and the diagonally
embedded S 1 is the maximal central torus Z in G = U(n). It is clear by inspection that
Z · T 0 = T , Z · G0 = G, and T 0 = T ∩ G0 , so ZG0 (T 0 ) = ZG0 (T ) = G0 ∩ ZG (T ) = G0 ∩ T = T 0 .
Thus, T 0 is a maximal torus in G0 , NG (T ) = Z · NG0 (T 0 ), and the inclusion Z ⊂ T implies
Z ∩ NG0 (T 0 ) = Z ∩ G0 = Z[n] ⊂ T 0 . Hence, the Weyl group W (G0 , T 0 ) is naturally identified
with W (G, T ) = Sn .
Under this identification of Weyl groups, the effect of Sn on the codimension-1 subtorus
T 0 ⊂ T = (S 1 )n via the identification of Sn with NG0 (T 0 )/T 0 is the restriction of the usual Sn action on (S 1 )n by permutation
of the factors. This is reminiscent of the representation of Sn
P
on the hyperplane { xj = 0} in Rn , and literally comes from this hyperplane representation
if we view T as (R/Z)n (and thereby write points of T as n-tuples (e2πixj )).
A key point is that, in contrast with U(n), in general the Weyl group quotient of NG0 (T 0 )
does not isomorphically lifts into NG0 (T 0 ) as a subgroup (equivalently, NG0 (T 0 ) is not a semidirect product of T 0 against a finite group). In particular, the Weyl group is generally only
a quotient of the normalizer of a maximal torus and cannot be found as a subgroup of the
normalizer. The failure of such lifting already occurs for n = 2:
Example 2.2. Let’s show that for G0 = SU(2) and T 0 the diagonal maximal torus, there is no
order-2 element in the nontrivial T 0 -coset of NG0 (T 0 ). A representative for this non-identity
coset is the “standard Weyl representative”
ν=
0 1
−1 0
3
that satisfies ν 2 = −1. For any t0 ∈ T 0 , the representative νt0 for the nontrivial class in the
order-2 group W (G0 , T 0 ) satisfies
−1
(νt0 )(νt0 ) = (νt0 ν −1 )(ν 2 t0 ) = t0 (−1)t0 = −1
regardless of the choice of t0 . Hence, no order-2 representative exists.
For n ≥ 3 the situation is more subtle, since (for G0 = SU(n) and its diagonal maximal
torus T 0 ) the order-2 elements of W (G0 , T 0 ) do always lift (non-uniquely) to order-2 elements
of NG0 (T 0 ). We will build such lifts and use them to show that the Weyl group doesn’t lift
isomorphically to a subgroup of NG0 (T 0 ). Inside W (G0 , T 0 ) = Sn acting in its usual manner
on the codimension-1 subtorus T 0 ⊂ T = (S 1 )n , consider the element σ = (12). This action
is the identity on the codimension-1 subtorus
T 00 := {(z2 , z2 , z3 , . . . , zn−1 , (z22 z3 · · · zn−1 )−1 ) ∈ (S 1 )n } ⊂ T 0
of T 0 (understood to consist of points (z2 , z2 , z2−2 ) when n = 3), so any g 0 ∈ NG0 (T 0 ) that lifts
σ ∈ W (G0 , T 0 ) must conjugate T 0 by means of σ and thus must centralize T 00 . Thus, the only
place we need to look for possible order-2 lifts of σ is inside ZG0 (T 00 ). But thinking in terms
of eigenlines once again, we see the action of T 00 on Cn decomposes as a plane Ce1 ⊕ Ce2
with action through z2 -scaling and the additional standard basis lines Cej on which T 00 acts
with pairwise distinct non-trivial characters.
Hence, the centralizer of T 00 inside the ambient group GLn (C) consists of elements that
(i) preserve the standard basis lines Cej for j > 2 and (ii) preserve the plane Ce1 ⊕ Ce2 .
This is GL2 (C) × (S 1 )n−2 . Intersecting with SU(n), we conclude that
ZG0 (T 00 ) = (SU(2) · S0 ) × S
where this SU(2) is built inside GL(Ce1 ⊕ Ce2 ), S is the codimension-1 subtorus of (S 1 )n−2
(in coordinates z3 , . . . , zn ) whose coordinates have product equal to 1, and
S0 = {(z, z, 1, . . . , 1, z −2 )}
(so SU(2) ∩ S0 = {±1}).
Clearly
NG0 (T 0 ) ∩ ZG0 (T 00 ) = (NSU(2) (T0 ) · S0 ) × S
where T0 ⊂ SU(2) is the diagonal torus. Thus, an order-2 element of NG0 (T 0 ) whose effect
on T 0 is induced by σ = (12) has the form (ν, s) for some 2-torsion ν ∈ NSU(2) (T0 ) · S0
acting on T0 through inversion and some 2-torsion s ∈ S. But ν cannot be trivial, and s
centralizes T 0 , so it is equivalent to search for order-2 lifts of σ subject to the hypothesis
s = 1. This amounts to the existence of an order-2 element ν ∈ NSU(2) (T0 ) · S0 that acts
on T0 through inversion. Writing ν = ν0 · s0 for s0 ∈ S0 and ν0 ∈ NSU(2) (T0 ), necessarily ν0
acts on T0 through inversion (so ν02 ∈ ZSU(2) (T0 ) = T0 ) and ν02 = s−2
0 ∈ S0 ∩ T0 = hτ i for
τ = (−1, −1, 1, . . . , 1).
A direct calculation in SU(2) shows that every element of NSU(2) (T0 ) − T0 has square
−2
1
equal to the unique element −1 ∈ T√
= ν02 = τ . This says
0 ' S with order 2, so s0
s0 = (i, i, 1, . . . , 1, −1) for some i = ± −1, so if n = 3 (so S = 1) then the order-2 lifts of
4
σ = (12) are exactly


0
it 0
−i/t 0 0 
0
0 −1
√
for some i = ± −1 and t ∈ S 1 . In particular, these all have trace equal to −1 when n = 3.
With a general n ≥ 3, assume we have a subgroup of NG0 (T 0 ) isomorphically lifting the
Weyl group. That identifies the lift with a subgroup Sn ⊂ NG0 (T 0 ) ⊂ G0 ⊂ GLn (C) acting
on the standard basis lines through the standard permutation because this subgroup has
conjugation effect on X(T = T 0 · ZG ) = Zn through the standard permutation. (We do not
claim the effect of this action permutes the set of standard basis vectors; it merely permutes
the set of lines they span.) In particular, n-cycles in Sn act by permuting the standard basis
lines and not fixing any of them, so such n-cycles act with trace equal to 0. For n = 3,
inspection of the characters of the 3-dimensional representation of S3 over C shows that
the only one with that trace property is the standard representation. But in the standard
representation of S3 the order-2 elements act with trace 1, and we saw above that if n = 3
then an order-2 lift in NG0 (T 0 ) of σ = (12) (if one exists) must have trace equal to −1. This
contradiction settles the case n = 3.
Finally, we can carry out an induction with n ≥ 4 (using that the cases of n − 1, n − 2 ≥ 2
are settled). Consider the subgroup Sn−1 ⊂ Sn stabilizing n. This acts on Cn preserving the standard basis line Cen , so its action on this line is one of the two 1-dimensional
representations of Sn−1 (trivial or sign). If trivial then we have
Sn−1 ⊂ SU(n − 1) ⊂ GL(Ce1 ⊕ · · · ⊕ Cen−1 )
normalzing the diagonal torus, contradicting the inductive hypothesis. Thus, this Sn−1 ⊂ Sn
acts through the sign character on Cen .
Now consider the subgroup Sn−2 ⊂ Sn−1 stabilizing 1, so its action on Cn preserves Ce1 .
The action of this subgroup on Cen is through the sign character (as the sign character on
Sn−1 restricts to that of Sn−2 ). If its action on Ce1 is also through the sign character then
by determinant considerations this Sn−2 is contained in
SU(n − 2) ⊂ GL(Ce2 ⊕ · · · ⊕ Cen−1 ),
contradicting the inductive hypothesis! Hence, its action on Ce1 is trivial. But (1n)conjuation inside Sn normalizes this subgroup Sn−2 and its effect on Cn swaps the lines
Ce1 and Cen (perhaps not carrying e1 to en !). Consequently, this conjugation action on
Sn−2 must swap the characters by which Sn−2 acts on Ce1 and Cen . But that is impossible
since the action on Ce1 is trivial and the action on Cen is the sign character. Hence, no lift
exists.
3. Character and cocharacter lattices
We want to show that W (G, S) := NG (S)/ZG (S) is finite for any torus S in G, and discuss
some applications of this finite group (especially for maximal S). The group W (G, S) with
its quotient topology is compact, and we will construct a realization of this topological group
inside a discrete group, so we can appeal to the fact that a discrete compact space is finite.
5
To analyze finiteness in a systematic way, it is useful to introduce a general concept that
will pervade the later part of the course.
Definition 3.1. The character lattice X(T ) of a torus T ' (S 1 )n is the abelian group
1
1 ⊕n
of (continuous) characters Hom(T, S 1 ) = Hom(S
= Zn (where (a1 , . . . , an ) ∈ Zn
Q ai , S )
corresponds to the character (t1 , . . . , tn ) 7→ ti ).
Underlying the description of X((S 1 )n ) as Zn is the fact that the power characters z 7→ z n
(n ∈ Z), are the endomorphisms of S 1 as a compact Lie group (as we saw in class earlier,
by using S 1 = R/Z). In an evident manner, T
X(T ) is a contravariant functor into the
category of finite free Z-modules, with X(T ) having rank equal to dim T .
Often the evaluation of a character χ ∈ X(T ) on an element t ∈ T is denoted in “exponential” form as tχ , and so correspondingly the group law on the character lattice (i.e.,
pointwise multiplication of S 1 -valued functions) is denoted additively: we write χ + χ0 to
denote the character t 7→ χ(t)χ0 (t), 0 denotes the trivial character t 7→ 1, and −χ denotes
the reciprocal character t 7→ 1/χ(t). This convention allows us to write things like
0
0
tχ+χ = tχ · tχ , t0 = 1, t−χ = 1/tχ .
The reason for preferring this exponential notation is due to the following observation. As
for any connected commutative Lie group (compact or not), the exponential map of a torus
is a canonical surjective homomorphism expT : t → T with discrete kernel Λ ⊂ t inducing
a canonical isomorphism t/Λ ' T . The compactness of T implies that Λ has maximal rank
inside the R-vector space t, which is to say that that natural map ΛR := R ⊗Z Λ → t is an
isomorphism. Thus,
X(T ) = Hom(t/Λ, S 1 ) = Hom(ΛR /Λ, S 1 ) ' Λ∗ ⊗Z Hom(R/Z, S 1 )
where Λ∗ denotes the Z-linear dual of Λ and the final isomorphism (right to left) carries ` ⊗ f
to the map R ⊗Z Λ → S 1 given by c ⊗ λ mod Λ 7→ f (c) · `(λ). But R/Z = S 1 via x 7→ e2πix ,
so plugging in the identification of the ring Hom(S 1 , S 1 ) with Z via power maps brings us
to the identification
Λ∗ ' Hom(t/Λ, S 1 ) = X(T )
given by
` 7→ (c ⊗ λ mod Λ 7→ e2πic`(λ) = e2πi`R (cλ) ).
Put another way, by identifying the scalar extension (Λ∗ )R with the R-linear dual of
ΛR = t we can rewrite our identification Λ∗ ' X(T ) as ` 7→ (v mod Λ 7→ e2πi`R (v) ) (v ∈ t).
In this manner, we “see” the character lattice X(T ) occuring inside analytic exponentials,
in accordance with which the S 1 -valued pointwise group law on X(T ) literally is addition
inside exponents.
Remark 3.2. A related functor of T which comes to mind is the covariant cocharacter lattice
X∗ (T ) = Hom(S 1 , T ) (finite free of rank dim T also). This is the Z-linear dual of the character
lattice via the evaluation pairing
h·, ·i : X(T ) × X∗ (T ) → End(S 1 ) = Z
defined by χ ◦ λ : z 7→ χ(λ(z)) = z hχ,λi (check this really is a perfect pairing, and note
that the order of appearance of the inputs into the pairing h·, ·i – character on the left and
6
cocharacter on the right – matches the order in which we write them in the composition
notation χ ◦ λ as functions).
We conclude from this perfect pairing that the covariant functor X∗ (T ) in T is naturally
identified with X(T )∗ = Λ∗∗ = Λ, and so its associated R-vector space X∗ (T )R is identified
with ΛR = t. Thus, we can expression the exponential parameterization of T in the canonical
form
expT : t/X∗ (T ) ' T.
The preference for X(T ) over X∗ (T ) will become apparent later in the course. To emphasize
the contravariance of X(T ) in T , one sometimes denotes the character lattice as X∗ (T ),
akin to the convention in topology that the contrvariant cohomology functors are denoted
with degree superscripts whereas the covariant homology functors are denoted with degree
subscripts. We will generally not do this.
Here is an elaborate triviality:
Lemma 3.3. The functor T
X(T ) from tori to finite free Z-modules is an equivalence of
categories. That is, for any tori T and T 0 and any map f : X(T 0 ) → X(T ) of Z-modules
there exists exactly one homomorphism φ : T → T 0 such that X(φ) = f , and every finite free
Z-module has the form X(T ) for a torus T .
In particular, passing to invertible endomorphisms, for an r-dimensional torus T the natural anti-homomorphism
Aut(T ) → GL(X(T )) ' GLr (Z)
is an anti-isomorphism.
Proof. It is clear that Zn ' X((S 1 )n ), so the main issue is bijectivity on Hom-sets. Decomposing T and T 0 into direct products of S 1 ’s and using the natural identification X(T1 ×T2 ) '
X(T1 ) × X(T2 ), we may immediately reduce to the case in which T, T 0 = S 1 . In this case
X(S 1 ) = Z via the power characters and the endomorphism ring of Z as a Z-module is
compatibly identified with Z.
4. Finiteness
Let’s return to tori in a connected compact Lie group G. For any such S in G, there
is a natural action of the abstract group W (G, S) = NG (S)/ZG (S) on S via n.s = nsn−1 ,
and by definition of ZG (S) this is a faithful action in the sense that the associated action
homomorphism W (G, S) → Aut(S) is injective.
The target group Aut(S) looks a bit abstract, but by bringing in the language of character
lattices it becomes more concrete as follows. Since the “character lattice” functor on tori is
fully faithful and contravariant, it follows that the associated action of W (G, S) on X(S) via
w.χ : s 7→ χ(w−1 .s) is also faithful: the homomorphism W (G, S) → GL(X(S)) ' GLr (Z) is
injective (with r = dim S = rankZ (X(S))).
Now we are finally in position to prove:
Proposition 4.1. The group W (G, S) is finite.
7
As we noted at the outset, this finiteness is equivalent to the equality ZG (S)0 = NG (S)0 .
(We cannot say ZG (S) = NG (S)0 until we know that ZG (S) is connected, which rests on the
Conjugacy Theorem.)
Proof. Since W (G, S) with its quotient topology is compact, it suffices to show that the
natural injective map W (G, S) → GL(X(S)) = GLr (Z) is continuous when the target is
given the discrete topology (as then the image of this injective map would be compact
and discrete, hence finite). Of course, it is equivalent (and more concrete) to say that
NG (S) → GL(X(S)) is continuous.
By using duality to pass from automorphisms of X(S) to automorphisms of X∗ (S), it is
equivalent to check that the injective homomorphism NG (S) → GL(X∗ (S)) arising from the
action of NG (S) on X∗ (S) defined by (n.λ)(z) = n · λ(z) · n−1 is continuous when the target
is viewed discretely. Equivalently, this says that the action map
NG (S) × X∗ (S) → X∗ (S)
is continuous when we give X∗ (S) the discrete topology and the left side the product topology.
But the discrete topology on X∗ (S) is the subspace topology from naturally sitting inside
the finite-dimensional R-vector space X∗ (S)R , so it is equivalent to check that the natural
action map
fS : NG (S) × X∗ (S)R → X∗ (S)R
is continuous relative to the natural manifold topologies on the source and target.
Recall that for any torus T whatsoever, the “realification” X∗ (T )R of the cocharacter
lattice is naturally identified by means of expT with the Lie algebra Lie(T ). Identifiying
X∗ (S)R with Lie(S) in this way, we can express fS as a natural map
NG (S) × Lie(S) → Lie(S).
What could this natural map be? By using the functoriality of the exponential map of Lie
groups with respect to the inclusion of S into G, and the definition of AdG in terms the effect
of conjugation on the tangent space at the identity point of G, one checks (do it!) that this
final map carries (n, v) to AdG (n)(v), where we note that AdG (n) acting on g preserves the
subspace Lie(S) since n-conjugation on G preserves S. (The displayed commutative diagram
just above (1.2) in Chapter IV gives a nice visualization for the automorphism ϕ = AdG (n)
of the torus S.)
So to summarize, our map of interest is induced by the restriction to NG (S) × Lie(S) of
the action map
G×g→g
given by (g, X) 7→ AdG (g)(X). The evident continuity of this latter map does the job.
Remark 4.2. What is “really going on” in the above proof is that since the automorphism
group of a torus is a “discrete” object, a continuous action on S by a connected Lie group
H would be a “connected family of automorphisms” of S (parameterized by H) containing
the identity automorphism and so would have to be the constant family: a trivial action on
S by every h ∈ H. That alone forces NG (S)0 , whatever it may be, to act trivially on S. In
other words, this forces NG (S)0 ⊂ ZG (S), so NG (S)0 = ZG (S)0 .
8
The reason that this short argument is not a rigorous proof as stated is that we have not
made a precise logical connection between the idea of “discreteness” for the automorphism
group of S and the connectedness of the topological space NG (S)0 ; e.g., we have not discussed the task of putting a natural topology on the set of automorphisms of a Lie group.
(The proof in IV, 1.5 that NG (T )0 = T is incomplete because of exactly this issue, so we
prove it below using Proposition 4.1, whose proof made this discrete/connected dichotomy
rigorous in some situations.) There is a systematic and useful way to justify this intuition
by means of the idea of “Lie group of automorphisms of a connected Lie group”, but it is
more instructive for later purposes to carry out the argument as above using the crutch of
the isomorphism X∗ (S)R ' Lie(S) rather than to digress into the topic (actually not very
hard) of equipping the automorphism group of a connected Lie group with a structure of
(possibly very disconnected) Lie group in a useful way.
Proposition 4.3. Let G be a compact connected Lie group, T a maximal torus. The identity
component NG (T )0 is equal to T (so NG (T )/T is finite).
Remark 4.4. The proof of the Conjugacy Theorem works with the finite group NG (T )/T ,
and it is essential in that proof that we know NG (T )/T is finite (but only after the proof
of the Conjugacy Theorem is over and connectedness of torus centralizers thereby becomes
available can we conclude that NG (T )/T = NG (T )/ZG (T ) =: W (G, T )).
Proof. To prove that NG (T )0 = T for any choice of maximal torus T in a compact connected
Lie group G, we first note that NG (T )0 = ZG (T )0 by Proposition 4.1. To show that ZG (T )0 =
T , if we rename ZG (T )0 with G (as we may certainly do!) then we find ourselves in the case
that T is central in G, and in such cases we wish to prove that G = T .
It suffices to show that the inclusion of Lie algebras t ,→ g is an equality. Choose X ∈ g and
consider the associated 1-parameter subgroup αX : R → G. Clearly αX (R) is a connected
commutative subgroup of G, so αX (R) · T is also a connected commutative subgroup of G
(as T is central in G). Its closure H is therefore a connected commutative compact Lie
group, but we know that the only such groups are tori. Thus, H is a torus in G, but by
design it contains T which is a maximal torus of G. Thus, H = T , so αX (R) ⊆ T . Hence,
0
αX : R → G factors through T , so X = αX
(0) ∈ t. This proves that g = t (so G = T ).
5. Subtori via character lattices
We finish this handout with a discussion of how to use the character lattice to keep track
of subtori and quotient tori. This is very convenient in practice, and clarifies the sense in
which the functor T
X(T ) is similar to duality for finite-dimensional vector spaces over a
field. (This is all subsumed by the general Pontryagin duality for locally compact Hausdorff
topological abelian groups, but our treatment of tori is self-contained without such extra
generalities.) As an application, we shall deduce by indirect means that any torus contains
a dense cyclic subgroup.
Let T be a torus. For any closed (not necessarily connected) subgroup H ⊂ T , the quotient
T /H is a compact connected commutative Lie group, so it is a torus. The pullback map
X(T /H) → X(T ) (composing characters with the quotient map T T /H) is an injection
9
between finitely generated Z-modules. For example, if T = S 1 and H = µn is the cyclic
subgroup of nth roots of unity (n > 0) then the map f : T → S 1 defined by f (t) = tn
identifies T /H with S 1 and the subgroup X(T /H) = X(S 1 ) = Z of X(T ) = X(S 1 ) = Z via
composition of characters with f is the injective map Z → Z defined by multiplication by
n. (Check this!)
Proposition 5.1. For any subtorus T 0 ⊂ T , the induced map X(T ) → X(T 0 ) is surjective
with kernel X(T /T 0 ). In particular, X(T /T 0 ) is a saturated subgroup of X(T ) in the sense
that the quotient X(T )/X(T /T 0 ) = X(T 0 ) is torsion-free.
Proof. The elements χ ∈ X(T ) killed by restriction to T 0 are precisely those χ : T → S 1 that
factor through T /T 0 , which is to say χ ∈ X(T /T 0 ) inside X(T ). This identifies the kernel, so
it remains to prove surjectivity. That is, we have to show that every character χ0 : T 0 → S 1
extends to a character T → S 1 (with all characters understood to be homomorphisms of Lie
groups, or equivalently continuous). We may assume T 0 6= 1.
Choose a product decomposition T 0 = (S 1 )m , so X(T 0 ) is generated as an abelian group by
the characters χ0i : (z1 , . . . , zn ) 7→ zi . (Check this!) It suffices to show that each χ0i extends to
a character of T . Let T 00 = (S 1 )m−1 be the product of the S 1 -factors of T 0 apart from the ith
factor. In an evident manner, χ0i is identified with a character of T 0 /T 00 = S 1 and it suffices
to extend this to a character of T /T 00 (as such an extension composed with T T /T 00 is an
extension of χ0i ). Thus, it suffices to treat the case T 0 = S 1 and to extend the identity map
S 1 → S 1 to a character of T .
Choosing a product decomposition T = (S 1 )n , the given abstract inclusion S 1 = T 0 ,→
T = (S 1 )n is a map ι : z 7→ (z e1 , . . . , z en ) for some integers e1 , . . . , en , and the
P triviality
of ker ι implies that gcd(ei ) = 1. Hence, there exist integers r1 , . . . , rn so that
rj ej = 1.
1
Hence, χ : T → S defined by
Y r
χ(z1 , . . . , zn ) =
zj j
satisfies χ(ι(z)) = z
P
rj ej
= z, so χ does the job.
Corollary 5.2. A map f : T 0 → T between tori is a subtorus inclusion (i.e., isomorphism
of f onto a closed subgroup of T ) if and only if X(f ) is surjective.
Proof. We have already seen that the restriction map to a subtorus induces a surjection
between character groups. It remains to show that if X(f ) is surjective then f is a subtorus
inclusion. For any homomorphism f : G0 → G between compact Lie groups, the image is
closed, even compact, and the surjective map G0 → f (G0 ) between Lie groups is a Lie group
isomorphism of G0 /(ker f ) onto f (G0 ). Thus, if ker f = 1 then f is an isomorphism onto a
closed Lie subgroup of G. Hence, in our setting with tori, the problem is to show ker f = 1.
By the surjectivity hypothesis, every χ0 ∈ X(T 0 ) has the form χ ◦ f for some χ ∈ X(T ),
so χ0 (ker f ) = 1. Hence, for any t0 ∈ ker f we have χ0 (t0 ) = 1 for all χ0 ∈ X(T 0 ). But by
choosing a product decomposition T 0 ' (S 1 )m it is clear that for any nontrivial t0 ∈ T 0 there
is some χ0 ∈ X(T 0 ) such that χ0 (t0 ) 6= 1 (e.g., take χ0 to be the projection from T 0 = (S 1 )m
onto an S 1 -factor for which t0 has a nontrivial component). Thus, we conclude that any
t0 ∈ ker f is equal to 1, which is to say that ker f = 1.
10
Corollary 5.3. The map T 0 7→ X(T /T 0 ) is a bijection from the set of subtori of T onto the
set of saturated subgroups of X(T ). Moreover, it is inclusion-reversing in both directions in
the sense that T 0 ⊆ T 00 inside T if and only if X(T /T 00 ) ⊆ X(T /T 0 ) inside X(T ).
Proof. Let j 0 : T 0 ,→ T and j 00 : T 00 ,→ T be subtori, so the associated maps X(j 0 ) and X(j 00 )
on character lattices are surjections from X(T ). Clearly if T 0 ⊆ T 00 inside T then we get a
quotient map T /T 0 T /T 00 as torus quotients of T , so passing to character lattices gives
that X(T /T 00 ) ⊆ X(T /T 0 ) inside X(T ).
Let’s now show the reverse implication: assuming X(T /T 00 ) ⊆ X(T /T 0 ) inside X(T ) we
claim that T 0 ⊆ T 00 inside T . Passing to the associated quotients of X(T ), we get a map
f : X(T 00 ) → X(T 0 ) respecting the surjections X(j 00 ) and X(j 0 ) onto each side from X(T ). By
the categorical equivalence in Lemma 3.3, the map f has the form X(φ) for a map of tori
φ : T 0 → T 00 , and j 00 ◦ φ = j 0 since we can check the equality at the level of character groups
(by how φ was constructed). This says exactly that T 0 ⊆ T 00 inside T , as desired.
As a special case, we conclude that X(T /T 00 ) = X(T /T 0 ) inside X(T ) if and only if T 0 = T 00
inside T . Thus, the saturated subgroup X(T /T 0 ) inside X(T ) determines the subtorus T 0
inside T .
It remains to show that if Λ ⊆ X(T ) is a saturated subgroup then Λ = X(T /T 0 ) for a
subtorus T 0 ⊆ T . By the definition of saturatedness, X(T )/Λ is a finite free Z-module, so
it has the form X(T 0 ) for an abstract torus T 0 . By the categorical equivalence in Lemma
3.3, the quotient map q : X(T ) X(T )/Λ = X(T 0 ) has the form X(φ) for a map of tori
φ : T 0 → T . By Corollary 5.2, φ is a subtorus inclusion since X(φ) = q is surjective, so T 0
is identified with a subtorus of T identifying the q with the restriction map on characters.
Hence, the kernel Λ of q is identified with X(T /T 0 ) inside X(T ) as desired.
Corollary 5.4. For a map f : T → S between tori, the kernel is connected (equivalently, is
a torus) if and only if X(f ) has torsion-free cokernel.
Proof. We may replace S with the torus f (T ), so f is surjective and hence S = T /(ker f ).
Thus, X(f ) is injective. If T 0 := ker f is a torus then
coker(X(f )) = X(T )/X(S) = X(T )/X(T /T 0 ) = X(T 0 )
is torsion-free. It remains to prove the converse.
Assume coker(X(f )) is torsion-free, so the image of X(S) in X(T ) is saturated. Thus, this
image is X(T /T 0 ) for a subtorus T 0 ⊂ T . The resulting composite map
X(f ) : X(S) X(T /T 0 ) ,→ X(T )
arises from a diagram of torus maps factoring f as a composition
T T /T 0 ,→ S
of the natural quotient map and a subtorus inclusion. It follows that ker f = T 0 is a torus. Since X(T ) has only countably many subgroups (let alone saturated ones), Corollary 5.3
implies that T contains only countably many proper subtori {T1 , T2 , . . . }. As an application,
we get the existence of (lots of) t ∈ T for which the cyclic subgroup hti generated by t is
dense, as follows. Pick any t ∈ T and let Z be the closure of hti (which we want to equal
T for suitable t), so Z 0 is a subtorus of T . This has finite index in Z, so tn ∈ Z 0 for some
11
n ≥ 1. We need to rule out the possibility that Z 0 6= T ; i.e., that Z 0 is a proper subtorus.
There are countably many proper subtori of T , say denoted T1 , T2 , . . . , and we just need to
avoid the possibility t ∈ [n]−1 (Tm ) for some n ≥ 1 and some m. Put another way, we just
have to show that T is not the union of the preimages [n]−1 (Tm ) for all n ≥ 1 and all m.
This is immediate via the Baire Category Theorem, but here is a more elementary approach.
If some tn (with n > 0) lies in a proper subtorus T 0 ⊂ T then tn is killed by the quotient
map T → T /T 0 whose target T /T 0 is a non-trivial torus (being commutative, connected,
and compact). Writing T /T 0 as a product of copies of S 1 and projecting onto one of those
d
factors, we get a quotient map q : T S 1 = R/Z killing tn . If we identify T with Rd /Z
, t is
P
d
d
d
represented by (t1 , . . . , td ) ∈ R and q : R /Z R/Z is induced by (x1 , . . . , xd ) 7→
aj x j
for some a1 , . . . , ad ∈ Z not all zero. Thus, we just have to pick (t1 ,P
. . . , td ) ∈ Rd avoiding the
possibility that for some n > 0 and some a1 , . P
. . , ad ∈ Z not all 0,
aj (ntj ) ∈ Z. Replacing
aj with naj , it suffices to find t ∈ Rn so that
bjP
tj 6∈ Z for all b1 , . . . , bdP∈ Z not all zero.
Suppose for some b1 , . . . , bd ∈ Z not all zero that bj tj = c ∈ Z. Then bj tj +(−c)·1 = 0,
so {t1 , . . . , td , 1} is Q-linearly dependent. Hence, picking a Q-linearly independent subset of
R of size dim(T ) + 1 (and scaling throughout so 1 occurs in the set) provides such a t. In
this way we see that there are very many such t.